Deep Learning Convolutional Neural Network Applications in Image Analysis to Identify Defects

Civil infrastructures like bridges, dams etc. must be regularly inspected for damage in the interest of public safety. Current inspections are done manually by human inspectors and are costly. More cost effective and accurate automated inspection technologies are needed. Drone imaging i.e. drones flying around buildings taking pictures and analysing them with computer image processing algorithms provides a possible technological solution. Many image processing techniques have been implemented before for defect detection on concrete structures, partially replacing human inspections. They manipulate images to extract defect features like cracks on concrete structures. Varying real-world situations like shadows, vegetation, dirt and dust over the inspected infrastructure, can pose a challenge for these techniques to accurately detect defects. To overcome such issues, this thesis proposes a Deep Convolutional Neural Networks (CNN) to detect cracks on concrete structures using images. The CNN is designed using transfer learning approach and trained on 40000 images containing both cracks and no-cracks. Care is taken so that the CNN does not overfit the training data using approaches like data augmentation and validation data tests. The trained CNN is also tested against images taken from drones and validated by human inspections. The final CNN achieved an accuracy of 99.71%.